3 ms·

The dual-actuator drives that Seagate is currently working on still have only one head per platter surface. They're not putting two sets of arms into the drive

The dual-actuator drives that Seagate is currently working on still have only one head per platter surface. They're not putting two sets of arms into the drive like the old Conner Chinook drives [1]. Instead, Seagate's dual-actuator drives mean that the arms for the top 3-4 platters move as one group, and the arms for the bottom 3-4 platters move as a second group. They're basically giving you two separate hard drives that share a spindle motor, and are in a transparent RAID-0. [1] https://en.wikipedia.org/wiki/File:Conner_Peripherals_%22Chinook%22_dual-actuator_drive.jpg https://en.wikipedia.org/wiki/File:Conner_Peripherals_%22Chi...

gruez 6y ago
What's the point of that? I guess it doubles random io performance? Not sure why that's even needed when there are SSDs.
- wtallis 6y ago
  IOPS/TB is an important metric for hard drives these days. Capacity increases have been pushing that number down, making the performance characteristics of larger hard drives incrementally more tape-like and less suitable for live data. Splitting the actuators gets hard drives back to the IOPS/TB they had a few generations ago, which makes them suitable for a slightly broader range of applications than they otherwise would be. It also means that an older system using an array of 8TB drives can be consolidated onto half as many 16TB drives with minimal other architecture changes. (Decreasing IOPS/TB is also why SATA SSDs haven't gone beyond 8TB while NVMe and SAS SSDs are pushing past 30TB.)
  - woofie11 6y ago
    I'd want to see an overall calculation. Twice as many actuators means twice as many things can go wrong. I think I prefer KISS. On the flip side, I guess each actuator gets used half as often. I use HDD for bulk storage and SSD for high-speed storage. I don't mind less speed, except insofar as it impacts reliability (e.g. a drive in a RAID fails, and the redundant drive fails while rebuilding the array). It seems like competitors to HDD have become increasingly noncompetitive over time. Tape drives are insanely expensive at reasonable capacity. DVDs hold 0.1% of an HDD and are more expensive. Various newfangled optical drives (e.g. M DISC) appear more reliable than HDDs, but also cost much more per TB. Plus, you need to keep a big pile of coasters.
    - hinkley 6y ago
      > I guess each actuator gets used half as often. I don't think that will be true. By having the heads only read half the platters, the bytes per track are <areal density increase>/2. That's more seek operations for linear reads and small offsets. For completely random access that's probably the same number of seek operations as before, but queuing latency is lower. Where it might reduce seeks is where multiple processes are trying to stream data at the same time, which sounds like is happening more and more.
  - notacoward 6y ago
    Dual actuators don't actually double IOPS/TB for all sorts of reasons - queuing effects, contention for resources elsewhere in the drive, or limitations of the external interface. In practice it tends to be far short of double, so you're better off planning for larger numbers of (possibly smaller) drives with the same capacity and IOPS than for single drives getting faster by any measure.
    - wtallis 6y ago
      Yeah, doubling IOPS/TB is just the theoretical limit. But dual-actuator drives will still be a much bigger performance improvement than NCQ was.
- cma 6y ago
  Should more than double it as each has less mass to move around.
- bityard 6y ago
  Raw performance isn't everything to everyone. Magnetic disks are still much cheaper than SSDs on a cost/GB basis and will continue to be for a good while yet. They are also more durable for many common workloads. As long as performance can be increased without increasing the cost too much, there will be a market for them.
- adrianmonk 6y ago
  Seagate's presentation made a big deal out of Total Cost of Ownership (TCO) and said most data centers still store 90% of data on hard drives and not SSD. Their goal is to retain that TCO edge by ensuring that HD performance doesn't get too much worse as HD capacity increases.

gruez 6y ago

What's the point of that? I guess it doubles random io performance? Not sure why that's even needed when there are SSDs.

wtallis 6y ago
IOPS/TB is an important metric for hard drives these days. Capacity increases have been pushing that number down, making the performance characteristics of larger hard drives incrementally more tape-like and less suitable for live data. Splitting the actuators gets hard drives back to the IOPS/TB they had a few generations ago, which makes them suitable for a slightly broader range of applications than they otherwise would be. It also means that an older system using an array of 8TB drives can be consolidated onto half as many 16TB drives with minimal other architecture changes. (Decreasing IOPS/TB is also why SATA SSDs haven't gone beyond 8TB while NVMe and SAS SSDs are pushing past 30TB.)
- woofie11 6y ago
  I'd want to see an overall calculation. Twice as many actuators means twice as many things can go wrong. I think I prefer KISS. On the flip side, I guess each actuator gets used half as often. I use HDD for bulk storage and SSD for high-speed storage. I don't mind less speed, except insofar as it impacts reliability (e.g. a drive in a RAID fails, and the redundant drive fails while rebuilding the array). It seems like competitors to HDD have become increasingly noncompetitive over time. Tape drives are insanely expensive at reasonable capacity. DVDs hold 0.1% of an HDD and are more expensive. Various newfangled optical drives (e.g. M DISC) appear more reliable than HDDs, but also cost much more per TB. Plus, you need to keep a big pile of coasters.
  - hinkley 6y ago
    > I guess each actuator gets used half as often. I don't think that will be true. By having the heads only read half the platters, the bytes per track are <areal density increase>/2. That's more seek operations for linear reads and small offsets. For completely random access that's probably the same number of seek operations as before, but queuing latency is lower. Where it might reduce seeks is where multiple processes are trying to stream data at the same time, which sounds like is happening more and more.
- notacoward 6y ago
  Dual actuators don't actually double IOPS/TB for all sorts of reasons - queuing effects, contention for resources elsewhere in the drive, or limitations of the external interface. In practice it tends to be far short of double, so you're better off planning for larger numbers of (possibly smaller) drives with the same capacity and IOPS than for single drives getting faster by any measure.
  - wtallis 6y ago
    Yeah, doubling IOPS/TB is just the theoretical limit. But dual-actuator drives will still be a much bigger performance improvement than NCQ was.
cma 6y ago
Should more than double it as each has less mass to move around.
bityard 6y ago
Raw performance isn't everything to everyone. Magnetic disks are still much cheaper than SSDs on a cost/GB basis and will continue to be for a good while yet. They are also more durable for many common workloads. As long as performance can be increased without increasing the cost too much, there will be a market for them.
adrianmonk 6y ago
Seagate's presentation made a big deal out of Total Cost of Ownership (TCO) and said most data centers still store 90% of data on hard drives and not SSD. Their goal is to retain that TCO edge by ensuring that HD performance doesn't get too much worse as HD capacity increases.